Abstract: Balancing content distribution between a machine learning model and a statistical model provides a baseline assurance in combination with the benefits of a well-trained machine learning model for content selection. In some implementations, a server receiving requests for a content item assigns a first proportion of the received requests to a first group and assigns remaining requests to a second group. The server uses a machine learning model to select variations of the requested content item for responding to requests assigned to the first group and uses a statistical model to select content variations for requests assigned to the second group. The server obtains performance information, e.g., acceptance rates for the different variations, and compares performance of the different models used for content selection. Audience share assigned to the machine learning model is increased when it outperforms the statistical model and decreased when it underperforms the statistical model.